Semiparametrically Efficient Estimation of Regression Models with Spillovers

Regression models with spillover effects generally cannot be estimated using ordinaryleast squares given the simultaneity that results from interactions among individuals.Instead, they are fitted using two-stage least squares (Kelejian and Prucha,1998; Bramoull´e et al., 2009), generalized method of moments (Liu et al., 2010), (quasi-)maximum likelihood typically under the normality assumption (Lee, 2004) or adaptiveestimation (Robinson, 2010).In this article, we propose a semiparametrically efficient estimator, based on theLocal Asymptotic Normality theory of Le Cam (1960) and on the work of Hallin et al.(2006, 2008) on residuals ranks-and-signs, that only requires strong unimodality of theerrors’ distribution as a distributional assumption. Monte Carlo simulations show thatthe suggested estimator performs well in comparison to competing estimators. A traderegression from Behrens et al. (2012) is used to illustrate how empirical findings mightgreatly change when the Gaussian distribution is not imposed.